def train_portfolio_helper(filename, pf_returns, mktrf, zero_init):
# drop missing returns
idx_missing = (pf_returns.iloc[:, 1] < -99)
pf_returns = pf_returns.loc[~idx_missing]
merged = pd.merge(pf_returns, mktrf, on='date')
# Retrieve portfolio info and print
pf_name = pf_returns.columns[1]
no_obs = len(merged)
no_lags = len(mktrf.columns) - 3
# return if the number of observations is not sufficient enough
if no_obs < (no_lags + 1) * 2:
return
print("")
print("*************************************************")
print("({})".format(datetime.today()))
print("portfolio: {}".format(pf_name))
print("no_obs: {}".format(no_obs))
print("*************************************************")
if (pf_name == 'smb') or (pf_name == 'hml'):
y_data = merged.loc[:, pf_name]
else:
y_data = merged.loc[:, pf_name] - merged.loc[:, 'rf']
y_data = y_data.as_matrix().reshape(-1, 1)
x_data = merged.iloc[:, 2:]
del x_data['rf']
x_data = x_data.as_matrix()
# Initialize the trainer
max_retries = 5
for attempt_id in range(max_retries):
trainer = Trainer(depth=2,
width=1,
no_inputs=x_data.shape[1],
zero_init=zero_init)
trainer.run_ols_regression(x_data, y_data)
params = trainer.train(x_data,
y_data,
x_tolerance=1e-6,
cost_tolerance=1e-6)
del trainer
if not check_if_overfitted_by_param(
params, freq="daily", no_lags=no_lags):
break
params = None
zero_init = False
print("Parameters are overfitted. Let's try again.")
print("")
# if parameters are still overfitted, replace parameters with OLS coefficients
if params is None:
trainer = Trainer(depth=2,
width=1,
no_inputs=x_data.shape[1],
zero_init=True)
trainer.run_ols_regression(x_data, y_data)
params = trainer.flush_params_to_dict()
# compute beta
if str.find(filename, 'daily') >= 0:
freq = 'daily'
else:
freq = 'monthly'
beta = compute_beta(param=params, freq=freq, no_lags=no_lags)
beta0 = beta[:, 0]
beta20 = np.sum(beta, axis=1)
beta_average = np.mean(beta20)
beta_delay = np.mean(beta20) - np.mean(beta0)
beta_convexity = (beta20[0] + beta20[-1]) / 2 - beta20[int(
(len(beta20) - 1) / 2)]
print("Results:")
print(" - filename: {}, portfolio: {}".format(filename, pf_name))
print(
" - beta_average: {:.3f}, beta_delay: {:.3f}, beta_convexity: {:.3f}".
format(beta_average, beta_delay, beta_convexity))
print("")
# save the results to SQL server if beta is not overfitted
if not check_if_overfitted_by_beta(beta):
sql_loader = DataLoader(connect=True)
query = """
select max(id)
from beta_portfolios
where filename = '{}' and portfolio = '{}' and lags = {}
""".format(filename, pf_name, no_lags)
max_id = sql_loader.sql_query_select(query)
if max_id.iloc[0, 0] is None:
next_id = 1
else:
next_id = max_id.iloc[0, 0] + 1
query = """
insert into beta_portfolios
VALUES
('{}', '{}', {}, {}, {}, now(), '{}', {}, {}, {})
""".format(filename, pf_name, no_lags, next_id, no_obs,
json.dumps(params), beta_average, beta_delay,
beta_convexity)
sql_loader.sql_query_commit(query)
sql_loader.close()
else:
print("* Do not save the results because the beta is overfitted *")
print("")
return
def train_stock_returns_full_periods(permno_from, permno_to):
no_lags = 10
print("***********************************")
print(" Train Stock Returns")
print(" permno_from: {}".format(permno_from))
print(" permno_to : {}".format(permno_to))
print(" no_lags : {}".format(no_lags))
print("***********************************")
loader = DataLoader(connect=True)
total_training_no = 0
total_training_time = 0
print("Loading market returns...")
mktrf = loader.load_market_returns('daily', no_lags=no_lags)
print("Loading untouched firm list...")
query = """
select permno, no_obs
from beta_stocks_full_periods
where permno >= {} and permno <= {} and touched is null
order by permno
""".format(permno_from, permno_to)
permno_list = loader.sql_query_select(query)
if permno_list is None:
print("Cannot find any untouched firm for permno between {} and {}".
format(permno_from, permno_to))
return
for i in range(len(permno_list)):
permno = permno_list.loc[i, 'permno']
print(" ")
print("*****************************************")
print(datetime.today())
print("permno: {} ({}/{})".format(permno, i + 1, len(permno_list)))
print("*****************************************")
t_start = time.time()
# load stock returns
print("Loading stock returns...")
stock_rets = loader.load_stock_returns(permno)
merged = pd.merge(mktrf, stock_rets, on='date')
no_obs = len(merged)
print("# observations: {}".format(no_obs))
if no_obs < 100:
print("Skip this stock due to the lack of observations (obs:{})".
format(no_obs))
continue
y_data = merged.loc[:, 'ret'] * 100 - merged.loc[:, 'rf']
y_data = y_data.as_matrix().reshape(-1, 1)
x_data = merged.iloc[:, 1:-1]
del x_data['rf']
x_data = x_data.as_matrix()
# Initialize the trainer
max_retries = 3
zero_init = True
for attempt_id in range(max_retries):
trainer = Trainer(depth=2,
width=1,
no_inputs=no_lags + 1,
zero_init=zero_init)
trainer.run_ols_regression(x_data, y_data)
params = trainer.train(x_data,
y_data,
x_tolerance=1e-2,
cost_tolerance=1e-3)
del trainer
if not check_if_overfitted_by_param(
params, freq="daily", no_lags=no_lags):
break
params = None
zero_init = False
print("Parameters are overfitted. Let's try again.")
print("")
# if parameters are still overfitted, replace parameters with OLS coefficients
if params is None:
trainer = Trainer(depth=2,
width=1,
no_inputs=no_lags + 1,
zero_init=True)
trainer.run_ols_regression(x_data, y_data)
params = trainer.flush_params_to_dict()
# compute beta
beta = compute_beta(param=params, freq='daily', no_lags=no_lags)
assert (not check_if_overfitted_by_beta(beta))
beta0 = beta[:, 0]
beta20 = np.sum(beta, axis=1)
beta_average = np.mean(beta20)
beta_delay = np.mean(beta20) - np.mean(beta0)
beta_convexity = (beta20[0] + beta20[-1]) / 2 - beta20[int(
(len(beta20) - 1) / 2)]
query = """
update beta_stocks_full_periods
set touched = now(),
parameters = '{}',
beta_average = {},
beta_delay = {},
beta_convexity = {}
where permno = {}
""".format(json.dumps(params), beta_average, beta_delay,
beta_convexity, permno)
loader.sql_query_commit(query)
elapsed = time.time() - t_start
total_training_no += 1
total_training_time += elapsed
print("Results:")
print(" - permno: {}, no_obs:{}".format(permno, no_obs))
print(
" - beta_average: {:.3f}, beta_delay: {:.3f}, beta_convexity: {:.3f}"
.format(beta_average, beta_delay, beta_convexity))
print("")
print("Elapsed time: {:.3f} seconds".format(elapsed))
print("")
print("Total training no: {}".format(total_training_no))
print("Total training time: {:.2f} minutes ({:.2f} hours)".format(
total_training_time / 60, total_training_time / 3600))
print("Average training time: {:.3f} seconds".format(
total_training_time / total_training_no))
loader.close()
